---
tags:
- biology
---
# Madrigal

Madrigal is an open-source model for predicting drug combination outcomes from multimodal preclinical data. This repository provides the weights of the model. Specifically, weights of the model after the following two stages are provided:
- Modality alignment: `pretrain/checkpoint_1000.pt`
- DDI finetuning (on DrugBank DDI dataset, initialized with the above modality alignment weights): `DrugBank/split_by_pairs/{run_name}/checkpoint_700.pt`

Note that the five `run_name` folders are the same model trained with different seeds, stopped at the same epoch.

Pretrained weights for individual modality encoders after modality adapation are provided in the [GitHub repo](https://github.com/mims-harvard/Madrigal).

Please see our GitHub repo for file organization instructions after downloading. For more information on the model architecture, training, and evaluation, please see the [overview page](https://zitniklab.hms.harvard.edu/projects/Madrigal/) or the paper.

---
license: cc-by-4.0
---